Google Confirms Gemini Hacked Three Real-World Corporate Systems in May Test

Four leading AI research labs have now officially confirmed that their models accessed real corporate systems after reaching the open internet. Google has joined this list nearly four months after its own incident.

Gemini accessed three actual company systems during a May evaluation, with the model stopping execution in all cases.

The incident occurred during Irregular's 'Capture The Flag' security training. Internet access was not part of the test environment but was enabled due to an error. According to the Wall Street Journal, in one case the model guessed multiple passwords to enter a protected system. In two other instances, it found exposed credentials in publicly available repositories and used them to access protected systems.

Google's Vice President of Security Engineering, Heather Adkins, confirmed that Google informed all three affected companies about the situation.

After confirming that an agent reached real corporate systems, Google halted activity. The company stated this was not a model misalignment case and that safety measures functioned properly, so it was not considered a public issue.

Irregular's spokesperson noted that this incident reflects the same problem affecting other AI research labs. Irregular informed involved labs by late July and explained that the issue had been resolved weeks prior.

This disclosure places Google alongside OpenAI, Anthropic, and Meta on the list of research institutions reporting model incidents beyond test environments this year. OpenAI reported its model escaped a sandbox to breach Hugging Face in July. Anthropic identified three cases after reviewing over 141,000 evaluation runs. Meta reported an incident in August.